2026
agentside — Open-source Agent Engine
Let an LLM operate any SaaS through its own API — safely, in production. MIT licensed.
TypeScriptLLM tool-useSKILL.mdMulti-providerMIT
The role
Creator & maintainer
Highlights
- Open-sourced the agent engine behind Airbase — MIT-licensed and small enough to read in an afternoon.
- Security by design: the agent never touches your database. Every action goes through your existing API as the logged-in user, so it can't escalate privileges it never had.
- Skills are SKILL.md files — the same convention Claude Code uses. Teach the agent a new task by writing markdown, no redeploy.
- Model-agnostic: sessions are stored and replayed, any LLM plugs in, and skills load lazily so the prompt stays small.
- Not a demo repo — this exact pattern serves real users in production.
The idea
While building Airbase I realized the agent backend wasn't specific to hiring at all — it's just an agent that operates an app, any app, through the app's own API. So I extracted it, cleaned it up, and MIT-licensed it as agentside.
Decisions I'd defend in a fight
- The agent never touches your database. Every action goes through your existing API, as the actual logged-in user, with the same permissions as always. The LLM can't escalate privileges because it never had any.
- No new capability format. Skills are
SKILL.mdfiles — the same convention Claude Code and other agents already use. If you've written one for Claude, you already know how to teach this agent your app. - Swappable models, replayable sessions, lazy-loaded skills. The prompt stays small; the system stays flexible.
It's running in production today — not a demo repo. If you're adding AI to your SaaS, it's free — take it.
Want something like this built?